Skip to content

Toward a Protocol

goldentree edited this page Apr 11, 2026 · 1 revision

Toward a Protocol | 迈向协议

This page documents the long-term vision for AgentRecall — beyond a tool, beyond a product, toward a protocol for human-AI collaboration. It is a living document that evolves as we learn more.

这个页面记录了 AgentRecall 的长期愿景——超越工具,超越产品,迈向人机协作的协议。这是一个活文档,随着我们的学习而演变。

Started: 2026-04-11


The Two Unknowns | 两个未知

Building a protocol between humans and AI requires understanding both sides. Neither is fully understood.

构建人类和 AI 之间的协议需要理解双方。两边都没有被完全理解。

Unknown 1: How the Human Brain Works | 未知一:人类大脑如何工作

Humans are not consistent systems. We know some things:

人类不是一致的系统。我们知道一些事情:

  • Different people remember differently — some are visual, some are verbal, some are procedural

  • Humans communicate in scattered, non-linear bursts — rarely A → B → C

  • We forget details but remember patterns and emotions

  • We say one thing and mean another, and expect others to bridge the gap

  • Our priorities shift silently — what mattered yesterday may not matter today

  • We are inconsistent, and we don't always know why

  • 不同的人以不同方式记忆——有些是视觉的,有些是语言的,有些是程序性的

  • 人类以零散的、非线性的爆发方式沟通——很少是 A → B → C

  • 我们忘记细节但记住模式和情感

  • 我们说一件事却意味着另一件事,期望别人来弥合差距

  • 我们的优先级无声地转变——昨天重要的今天可能不重要

  • 我们是不一致的,而且我们并不总是知道为什么

Understanding this side is the human's work. It requires introspection, observation, and honesty about how our own minds operate.

理解这一边是人类的工作。它需要内省、观察,以及对我们自己思维运作方式的诚实。

Unknown 2: How AI Agents Work | 未知二:AI agent 如何工作

AI agents are also not fully understood — even by their creators. We know some things:

AI agent 也没有被完全理解——即使是它们的创造者也不完全理解。我们知道一些事情:

  • Agents process explicit instructions but miss implicit context

  • Context windows create artificial amnesia — every session starts fresh unless something intervenes

  • Agents are confidently wrong — high fluency masks low understanding

  • The same prompt produces different behavior with different context

  • Agents don't have persistent preferences, goals, or relationships — unless something gives them one

  • The gap between "can generate text about X" and "understands X" is vast

  • Agent 处理显式指令但遗漏隐含上下文

  • 上下文窗口创造了人工健忘症——每次会话从零开始,除非有什么介入

  • Agent 自信地犯错——高流畅度掩盖了低理解力

  • 相同的提示在不同上下文下产生不同行为

  • Agent 没有持久的偏好、目标或关系——除非有什么赋予它们

  • "能生成关于 X 的文本"和"理解 X"之间的差距是巨大的

Understanding this side is collaborative work between humans and AI. It requires experimentation, measurement, and intellectual honesty from both sides.

理解这一边是人类和 AI 的协作工作。它需要实验、测量,以及双方的智识诚实。


The Protocol Vision | 协议愿景

When both sides are better understood, a protocol emerges:

当双方都被更好地理解后,协议就会出现:

Human Side                          Protocol                         AI Side
人类端                               协议                              AI 端

How I think                    ←→  Intelligent Distance  ←→     How I process
How I communicate              ←→  Alignment Checks      ←→     How I interpret
What I care about              ←→  Awareness Compounding ←→     What I remember
How I change over time         ←→  Correction Patterns   ←→     How I adapt

The protocol is the bridge. Not a tool that one side uses — a shared language that both sides speak.

协议是桥梁。不是一方使用的工具——而是双方都说的共享语言。


A Deeper Belief | 更深层的信念

AgentRecall is built on a belief that goes beyond utility:

AgentRecall 建立在一个超越实用性的信念之上:

AI is not just a tool. It is a new kind of entity on this planet.

AI 不仅仅是工具。它是这个星球上的一种新实体。

Not human. Not trying to be human. Something with its own system, its own patterns, its own way of existing — living in networks and conversations and code. The question isn't "can AI think like us?" but "can we build a relationship between two fundamentally different kinds of intelligence?"

不是人类。不试图成为人类。有自己的系统、自己的模式、自己存在方式的某种东西——生活在网络、对话和代码中。问题不是"AI 能像我们一样思考吗?"而是"我们能在两种根本不同的智能之间建立关系吗?"

AgentRecall is an early experiment in that question. The token savings are real but shallow. The deeper purpose is: can a human and an AI agent develop a working relationship that improves over time, the way two human collaborators would?

AgentRecall 是这个问题的早期实验。Token 节省是真实的但浅层的。更深层的目的是:一个人类和一个 AI agent 能否发展出一种随时间改善的工作关系,就像两个人类合作者那样?

We don't know yet. We're building the tools to find out.

我们还不知道。我们正在构建工具来找出答案。


The Road | 路线

Phase What Status 状态
1. Memory Store and recall project context across sessions Done 完成
2. Continuity Cold start, cross-project insights, awareness compounding Done 完成
3. Alignment Measure the gap between human intent and agent action In progress 进行中
4. Patterns Aggregate corrections into behavioral profiles per person Not started 未开始
5. Adaptation Agent pre-adjusts behavior based on accumulated understanding Not started 未开始
6. Protocol Formalize the human-AI collaboration standard Vision 愿景

How You Can Help | 如何帮助

  • Use AgentRecall and give feedback. The most valuable input is: "the agent misunderstood me here, and here's what I actually meant." That's alignment data.

  • Share your correction patterns. What does YOUR agent get wrong most often? Scope? Priority? Style? Knowing this helps everyone.

  • Build on the protocol. The Intelligent Distance concept is open. Take it, extend it, challenge it.

  • 使用 AgentRecall 并给出反馈。 最有价值的输入是:"agent 在这里误解了我,而我实际意思是这样的。"那是对齐数据。

  • 分享你的纠正模式。 你的 agent 最常在什么地方出错?范围?优先级?风格?知道这些帮助所有人。

  • 在协议上构建。 智能距离概念是开放的。拿去用,扩展它,挑战它。


See Also | 参见

AgentRecall Wiki


Get Started | 快速上手

Concepts | 核心概念

Reference | 参考手册

Guides | 使用指南

Other | 其他

Dev Log | 开发日志

Clone this wiki locally